Application of artificial intelligence in orthopaedic research: From preclinical to translational

Abstract

Artificial intelligence (AI) has become pervasive in biomedicine and is transforming orthopaedic research from bench to bedside. Beyond its established roles in robotic surgery and diagnostics, AI now supports advances in biomechanics, imaging, tissue engineering, drug discovery, genomics, and prosthetic control. In biomechanics, AI enables faster finite-element simulations, markerless gait analysis, and data augmentation using synthetic signals. Imaging applications include automated segmentation of the spine and hip, opportunistic screening for osteoporosis, bone metastasis detection, and three-dimensional analysis of knee osteoarthritis. In regenerative medicine, AI assists in scaffold optimization, bioprinting, and personalized cell therapies, while integration with genomic and proteomic data enhances precision orthopaedics. Machine learning–based control systems also improve the usability of prosthetics and exoskeletons, reducing cognitive burden for patients. Despite challenges such as data scarcity, validation, and ethical considerations, AI is emerging as a powerful complement to traditional research methods. By accelerating workflows, improving accuracy, and enabling individualized care, AI holds strong potential to bridge laboratory discoveries with clinical applications in orthopaedics. This review highlights the application of AI in orthopaedic research and assesses how it could integrate into clinical practice in the future.

Introduction

Artificial intelligence (AI) has emerged as a transformative force across multiple disciplines, and orthopaedic surgery is no exception. Beyond its established roles in robotic surgery, diagnostic support, and treatment optimization, AI is now reshaping the landscape of preclinical and translational orthopaedic research, and tools to enhance diagnostic accuracy have been employed in clinical orthopaedics.

Traditional preclinical methods in orthopaedics are often time-consuming, resource-intensive, and limited in predictive power. AI has begun to bridge the gap by enhancing animal model analysis, biomarker discovery, drug and device development, and biomechanical simulation. ¹ In particular, machine learning (ML) and deep learning (DL) techniques have facilitated more accurate interpretation of laboratory data, enabling faster translation of experimental findings into clinical applications. ¹

In biomechanics and computational modelling, AI has helped through detailed simulations of the mechanics of prostheses and implants. ² In tissue engineering and regenerative medicine, AI supported the scaffold design through the biological and mechanical properties for bone and cartilage regeneration performance. ³ AI has also enhanced drug discovery and its delivery systems in orthopaedic research and care. ⁴ Integration with genomic datasets is further paving the way for personalized medicine, tailoring treatments based on individual biological profiles. ⁵

This review highlights the expanding role of AI in advancing preclinical and translational orthopaedic research, focusing on its applications in imaging, biomechanical analysis, tissue engineering, and drug/device development.

AI-assisted imaging and analysis in preclinical orthopaedic research

AI in biomechanical analysis

AI and finite element analysis (FEA)

Finite Element Analysis is widely used in orthopaedic biomechanics but is computationally demanding (Table 1). Recent developments show that AI can generate real-time FEA outputs, enabling iterative implant design and intraoperative decision-making. Hashemi et al. introduced a machine learning–based surrogate modelling framework that accelerates computational biomechanics by training models on large FEA-generated datasets. ⁶ These surrogate models provide rapid yet reliable predictions of stress–strain responses and musculoskeletal mechanics, dramatically reducing computation time while maintaining accuracy. This approach holds promise for prosthetic and implant design, rehabilitation planning, and personalized biomechanics, where speed and precision are critical.

Table 1.

AI in biomechanical analysis - AI in finite element analysis (FEA).

Reference	Methods	AI methods	Key findings	Notes
Hashemi et al., 2023	Finite element analysis (FEA) outputs used as training data.	Surrogate modeling using machine learning algorithms (specific models: e.g., regression-based, neural networks; trained on FEA data).	ML surrogates reproduced FEA outputs with high accuracy (low prediction error reported) and dramatically reduced computational time compared to full FEA.	Generalizability limited; external validation across populations and conditions still needed; model trained on simulation data only (not direct clinical data).

AI in gait and movement analysis

AI techniques such as Support Vector Machines (SVMs), Artificial Neural Networks (ANNs), and k-nearest neighbors (k-NNs) have been increasingly applied to gait biomechanics (Table 2). While marker-based 3D motion capture remains the gold standard, its high cost and technical requirements limit clinical utility. Recent advances in markerless AI-based video systems (Ino et al., Sipari et al.) demonstrate comparable accuracy using more accessible, non-invasive approaches.^7,8

Table 2.

AI in biomechanical analysis – AI in gait and movement detection.

Reference	Method/Device	AI method	Main outcomes	Accuracy/Performance
Lim et al. (2022)	IMUs	SVM, RF, NN	AI is effective in pathological vs. healthy gait.	Up to >90%
Ben Chaabane et al. (2023)	3D motion capture	LSTM, MLP, CNN, ResNet, ViT	Prediction of gait progression via GPS.	AUC >0.72
Ino et al. (2023)	Single camera video (OpenPose)	Pose estimation AI	Valid for bilateral limb analysis.	MAE 2.3–4.1°, CMC 0.89–0.99
Molavian et al. (2023)	Various (IMU, cameras)	ANN, SVM, k-NN	AI is useful for biomechanics & sports.	Not quantitative (review)
Sharifi renani et al. (2021)	IMUs + synthetic data	Neural networks	Synthetic data improves kinematics prediction.	RMSE ↓ to 1.7–1.9°
Sipari et al. (2022)	SANE markerless AI	PoseNet, OpenPose	Low-cost AI gait analysis.	Comparable to lab MoCap
Tedesco et al. (2020)	IMUs	MLP, Gradient boosting, SVM	Differentiate post-ACL gait.	Accuracy 73%, sens. 81.8%

Systematic reviews confirm that AI models outperform conventional methods in analyzing high-dimensional gait datasets. ⁹ Both studies, Lim et al. and Molavian et al., confirmed the diagnostic utility of AI in distinguishing pathological gait, particularly in neurological conditions like Parkinson’s disease, with high accuracy (>90%).^9,10 From the sports applications, Tedesco et al. demonstrated that IMU-based ML models detected long-term gait alterations with ACL injury, even years after the initial trauma. ¹¹

Aside from the diagnostic approach, Ben Chaabane et al. showed that advanced AI architectures (LSTM, CNN, ResNet, Vision Transformers) can forecast gait progression, supporting individualized treatment planning. ¹² One of the main challenges in applying AI to gait research is the scarcity of data. To overcome this, Sharifi Renani and colleagues ¹³ developed synthetic IMU signals for model training, which significantly improved prediction accuracy, reducing errors by more than half at the hip and nearly as much at the knee. While these advances are promising, limitations remain. Much of the work so far has been carried out in controlled laboratory settings, making real-world validation necessary before clinical use. In addition, approaches such as unsupervised and transfer learning are still underutilized, even though they could enhance robustness in natural environments. Better integration of multimodal inputs—such as video, IMU, and EMG—would also provide a more comprehensive view of movement. Finally, ethical and privacy issues surrounding gait data must be carefully addressed before these technologies can be adopted more broadly.^8,11

AI-assisted pre-clinical imaging

3D Micro-CT segmentation and growth-plate detection

In this subsegment, we highlighted four articles that discussed the use of AI for CT-segmentation in mice (Table 3). While Lagzouli et al. developed an AI model to segment cortical and trabecular bone in 3D high-resolution µCT scans of mouse tibia, ¹⁴ Burlutsky et al. developed a public dataset of 83 µCT mouse femur scans with growth plate annotations. ¹⁵ Lagzouli et al. presented the Dual-Branch Attention-based Hybrid Network (DBAHNet). This hierarchical architecture combines transformers and convolutional neural networks (CNNs) to segment cortical and trabecular regions in 3D µCT scans of the mouse tibia. Results from Lagzouli et al.’s study demonstrated that DBAHNet can provide a reliable and accurate 3D mouse µCT segmentation with Dice similarity coefficients (DSC) of 98.41% overall, 99.13% for cortical bone, and 97.69% for trabecular bone. ¹⁴ Burlutsky et al. have developed a six-version deep learning method to identify the location of the bone growth plane as a key step for automatic trabecular bone segmentation. These deep learning models achieved a mean absolute error of 1.91 ± 0.87, indicating that the data accuracy is within an acceptable level. ¹⁵

Table 3.

AI – assisted pre-clinical imaging - 3D Micro-CT segmentation and growth-plate detection.

Reference	AI methodology & data	Key findings	Recommendation
Lagzouli, A. et al. (2025)	Dual-branch attention-based hybrid network (DBAHNet) combining CNNs and transformers for 3D high-resolution µCT segmentation.	DBAHNet achieved state-of-the-art performance in segmenting cortical and trabecular compartments in mouse tibiae. It achieved dice similarity coefficients of up to 98.41% and demonstrated robustness across diverse conditions (drug treatments, mechanical loading, varying resolutions, and mouse strains).	DBAHNet can automate µCT bone segmentation in preclinical studies, reducing manual workload and improving reproducibility. Recommended for integration into automated pipelines for bone remodeling and drug-testing research.
Burlutskiy et al. (2024)	Six deep learning approaches (3D CNN regression, 2.5D long-axis regression, 2D/2.5D axial classification) tested in MiceBoneChallenge.	Mean absolute error (MAE) of growth plate detection: 1.91±0.87 planes from ground truth. Public dataset of 83 µCT scans and annotated growth plates released.	Public datasets and benchmark solutions should be used to develop and compare AI models for bone quantification in preclinical studies.

AI – Preclinical screening and disease modelling, assisted sports injury prediction

Machine learning based osteoporosis model

In the Preclinical model, machine learning or AI assists in various ways (Table 4). Yang et al. have developed a machine learning model to generate a screening tool for osteoporosis risk based on the Hong Kong Chinese population. ¹⁶ This model used a Naïve Bayes algorithm with seven easily measurable predictors (age, gender, education level, decreased height, BMI, teeth lost, vitamin D intake). The Preclinical Osteoporosis Screening Tool (POST) achieved high accuracy (AUC 0.86; sensitivity and specificity 83%), even outperforming the traditional statistical model, suggesting a promising future for POST. ¹⁶ In Paek et al. study, machine learning tried to replicate bone physiology by integrating a CNN with a biomimetic bone-on-a-chip platform. ¹⁷ The system recapitulated the osteon microenvironment using osteoblast-derived ECM and cocultured osteocytes/osteoblasts, enabling accurate replication of bone physiology. This model provides a uniform, transparent imaging window that enables the observer to observe cell-cell interactions and supports high-throughput bone unit analysis. The model is also compatible with a high-content screening system that enables drug tests, such as an anti-SOST antibody drug for osteoporosis. ¹⁷

Table 4.

Machine learning based on an osteoporosis model.

Reference	AI methodology & data	Key findings	Recommendation
Paek K et al. (2023)	AI-based deep learning (CNN with convolution + pooling layers) for image analysis of β-catenin nuclear translocation in bone-on-a-chip assays.	Developed a high-throughput biomimetic bone-on-a-chip that mimics osteon structure with osteoblast/osteocyte co-culture. AI-assisted analysis successfully evaluated the efficacy of an anti-SOST antibody (osteoporosis drug) by detecting β-catenin translocation.	An AI-assisted bone-on-a-chip is recommended as an efficient preclinical drug-testing platform for osteoporosis. It offers high-throughput, reproducibility, and translational potential for bone-related drug development.
Yang et al. (2023)	Machine learning–based preclinical osteoporosis screening tool (POST). Models tested: Gradient boosting machine, support vector machine, naïve bayes, and logistic regression. The best-performing simplified model was Naïve bayes.	Identified seven accessible predictors (age, gender, education, decreased body height, BMI, teeth lost, vitamin D supplement intake). POST achieved AUC 0.86, sensitivity 83%, specificity 83%, and accuracy 83%. Balanced high sensitivity and specificity compared to existing tools.	POST is a simple, cost-effective, and accurate screening tool for preclinical osteoporosis risk. Recommended for population-level screening and guiding targeted DXA testing in clinical practice.

Sport injury prediction

AI is increasingly used to detect biomechanical risk signatures for sports injuries by combining wearable sensor kinematics, strength, and stiffness measures, and laboratory-based references (Table 5). Recent studies have demonstrated strong in-sample classification for predicting anterior cruciate ligament (ACL) injury risk using inertial sensors and support vector machines (SVMs). According to Taborri et al., the important ACL risk indicators include stability, impact absorption, and joint excursion. ¹⁸ Other than ACL risk indicators, robotic pivot-shift experiments quantified ATT and ETR with high reproducibility, clearly distinguishing ACL-deficient from intact knees. Such benchmarks may serve as gold-standard labels for training AI models that estimate instability in wearable or video data. Robotic pivot-shift testing provides reproducible, high-fidelity kinematic profiles of ACL instability. ¹⁹ Integrating such data with wearable sensor outputs could enhance AI model training and reduce reliance on subjective evaluations.

Table 5.

AI in sports injury prediction.

Reference	AI methodology & data	Key findings	Recommendation
Amendolara A et al. (2023)	Narrative/mini-review of ML for sports-injury prediction (2017–2022). Algorithms discussed include K-NN, K-means, decision trees, random forests, gradient boosting/AdaBoost, and neural networks.	Concludes there are no strong, real-world efficacy conclusions yet due to a lack of uniform open-source datasets and reliance on older regression methods; addressing these gaps is needed to deploy powerful, modern ML.	The review synthesizes multiple sports (soccer, futsal, hockey, CrossFit, etc.), highlighting emerging work but noting inconsistent data/reporting standards across studies.
Colbrunn RW et al. (2019)	Robotic musculoskeletal simulator that reproduces the clinical pivot-shift exam using kinetic loading profiles with real-time force-feedback control (no ML).	Identified a loading profile that induced a positive pivot shift in all ACL-deficient specimens (p < 0.001). Shifts began at 22 ± 8° and ended at 33 ± 6°of knee flexion; magnitude 12.8 ± 3.2 mm anterior tibial translation and 17.6 ± 4.3° external tibial rotation.	The approach aims to mimic the dynamic clinical exam rather than static rotary loads; profiles were scaled/validated against surgeon-measured kinematics and then tested across multiple cadaveric knees.
Calderón-Díaz M et al. (2024)	Explainable ML for soccer hamstring-injury classification. Trained 35 classifiers (DT, discriminant, logistic regression, NB, SVM, KNN, ensembles, boosted/bagged trees, ANNs, XGBoost) on 110 male pros with 19 biomechanical/anthropometric features.	XGBoost reached a precision of up to 78%; hamstrings’ maximum strength and stiffness emerged as the most differentiating predictors.	Dataset built from Chilean first-division teams; tests included nordic hamstring, quadriceps chair, single-leg bridge, myotonometry, and bosco jump battery; preprocessing used scaling and zero-imputation; 10-fold CV.
Taborri J et al. (2021)	Supervised ML to flag female basketball players at risk for non-contact ACL injury using an IMU + optoelectronic bars; 9 classifiers were evaluated. Linear SVM performed best.	Best model (linear SVM) achieved an accuracy 96% and F1-score 95%; features with the most discriminative power included ellipse area, load absorption, leg mobility; ellipse area strongly correlated with the LESS score.	LESS was used as the supervisory clinical label; the protocol included jump-landing for LESS, plus monopodalic countermovement jumps (mCMJ) and single-leg squats with IMU/opto bars.

Besides ACL injury risk, muscle injury in soccer can be predicted by identifying hamstring strength and stiffness. In muscle-injury prediction, XGBoost classifiers applied to professional soccer players identified maximal hamstring force and tissue stiffness as top features, achieving ∼78% accuracy. ²⁰ Importantly, this approach provided interpretable outputs, underscoring the need for explainability in clinical translation. However, a study by Amendolara et al. found that the current lack of open-source datasets and reliance on dated regression models produce inconclusive results regarding the efficacy of machine learning in predicting sports injuries. ²²

AI can classify athletes into higher- and lower-injury-risk strata based on movement stability and load-absorption metrics. In youth basketball players, inertial-sensor features such as ellipse area and RMS acceleration during landing correlated with LESS scores, producing high accuracy (96%) with a linear SVM. ¹⁸

Tissue engineering and regenerative medicine

AI is slowly becoming an indispensable tool in orthopaedic research, where it can extract meaningful patterns and insights from large, heterogeneous datasets, thereby improving prediction and guiding experimental design. Applications include bone imaging analysis, surgical planning, and regenerative orthopaedics. For example, Fredrik Liu et al. developed a neural network model that integrated preclinical and clinical data to predict cartilage repair outcomes, achieving a cross-validated R² = 0.637 ± 0.005 (p < 0.005). ²³ The model identified that the defect depth, area, and implantation dose were essential predictors, with optimal MSC doses ranging between ∼17 and 25 million cells. Importantly, it also quantified uncertainty, providing confidence-based guidance for clinical translation.

Many studies have highlighted how AI can assist in musculoskeletal imaging, perioperative planning, and regenerative applications; however, they also mention significant limitations in data quality, generalizability, and transparency.^21–25 AI’s role in regenerative orthopaedics is most evident in scaffold design, bioprinting, and optimization of cell therapies (Table 6). Wang et al. used machine learning (ML) to generate spinodal microstructures, enabling fine-tuning of porosity and mechanics. Research by Conev et al. demonstrates that ML-guided 3D printing can support or enhance scaffold-fidelity optimization. ²⁶ A study by Malekpour & Chen (2022) showed integrated computational and ML models can help in bioink printability and cell viability, addressing a major issue in extrusion bioprinting. ²⁷

Table 6.

Tissue engineering and regenerative medicine in orthopaedic research.

Reference	Results	Key contributions
Kumar et al., 2025	Reviewed 59 studies on AI in orthopaedic care, with a dedicated section on bone healing and regeneration. Discussed AI’s role in diagnostics, intraoperative robotics, and postoperative recovery, while specifically detailing how AI enhances bone tissue regeneration devices, neuroprosthetics, and biomaterials optimization. Reported measurable outcomes such as TiRobot reducing radiation exposure by ∼70% and increasing surgical efficiency by ∼20%, as well as Mako providing sub-millimeter implant placement accuracy.	Positioned AI as a multifunctional enabler across orthopaedic and regenerative workflows. Key contribution was linking robotics, imaging, and biomaterials informatics into a continuous orthopaedic TE pipeline. Provided evidence-based insights into how AI shortens the scaffold/implant development cycle and improves clinical outcomes in regenerative settings.
Lyu et al., 2025	Proposed an intelligent manufacturing platform for osteoarthritis organoids, using AI-driven algorithms at multiple levels: CNNs for microscopic image quality assessment, bayesian optimization for fabrication parameters, reinforcement learning (RL) for adaptive manufacturing control, and GANs for generating synthetic data to augment limited datasets. Demonstrated significant improvements in organoid morphology uniformity, cell viability, and reproducibility compared to conventional manual approaches.	Contributed a closed-loop “sense–decide–act” framework for OA organoid fabrication. Highlighted how AI allows continuous learning during biomanufacturing, enabling standardization, reproducibility, and scaling of complex tissue models — an essential step for translational tissue engineering in musculoskeletal disease.
Madhan Jeyaraman et al., 2023	Provided a broad review of AI and ML applications in regenerative orthopaedics, mapping applications from scaffold design, drug discovery, and predictive genomics to rehabilitation support. Notably discussed KeyGenes (for stem-cell differentiation profiling) and CellNet (for fate mapping), as well as DL-based algorithms applied to scaffold biodegradation datasets. Emphasized that while preclinical applications are promising, clinical deployment is still rare, citing the lack of validated models and ethical safeguards as barriers.	Key contribution was synthesizing the regenerative orthopaedics landscape, showing how AI integrates across multiple modalities (cell, scaffold, and clinical outcomes)—emphasized the dual role of AI as both a scientific discovery accelerator (e.g., scaffold modeling) and a decision-support tool for clinicians, while outlining validation and ethics as critical bottleneck.
Kolomenskaya et al., 2024	Systematic review of AI in bone tissue engineering. Covered applications in implant design, scaffold fabrication, biomechanical simulations, and in vivo integration monitoring. Specific use cases included CNN-based scaffold porosity prediction, GAN-driven scaffold generation, and RL-enhanced fabrication processes. Highlighted that successful AI adoption depends on large, high-quality datasets and rigorous clinical trials to ensure reliability and reproducibility.	Provided a stage-by-stage roadmap for AI in BTE, identifying how algorithms align with each step (design, printability, testing, monitoring). A crucial contribution was bridging computational scaffold optimization with real-world manufacturing, emphasizing the translational challenge of moving from in silico models to clinical-grade implants.
Farhadi et al., 2022	Broad review of AI in orthopaedic surgery, with subsections on classification, segmentation, detection, robotics, natural language processing (NLP), and outcome prediction. While primarily surgical, the paper contextualized AI’s indirect impact on TE, e.g., imaging segmentation for scaffold evaluation, predictive modeling for patient outcomes, and robotics for precision placement of implants. Limitations identified include data bias, lack of generalizability, and regulatory barriers.	Positioned regenerative TE within the larger orthopaedic AI ecosystem, stressing that translational TE solutions must integrate seamlessly with existing surgical and imaging AI frameworks. The contribution was to provide a macro-level view of AI readiness and clarify adoption barriers that TE researchers must address to align with surgical practice.
Vaishya et al., 2024	PRISMA-based scoping review of 18 studies (2013–2024) on AI in regenerative orthopaedics. Documented applications in stem cell therapy, platelet-rich plasma (PRP), scaffold biomaterials, and tissue engineering constructs. Discussed AI’s role in personalized regenerative medicine by tailoring treatment to patient-specific data, highlighted ethical issues, and the scarcity of large validation studies as barriers to clinical adoption.	Key contribution was a systematic evidence synthesis showing that AI is already embedded across regenerative workflows but remains fragmented and under-validated. Positioned AI as a potential “game changer” while warning that ethics and reproducibility are essential before clinical translation.
Wang et al., 2023	Developed a machine-learning framework to design and screen spinodal scaffold microarchitectures. Used a 3D CNN generator and property predictor to rapidly generate thousands of scaffold designs with controlled anisotropy and graded porosity. Validated results by predicting and achieving target elastic properties in silico, which would otherwise require computationally expensive phase-field simulations.	The first study in this collection provides a high-throughput scaffold design platform that significantly accelerates the discovery of scaffolds with desired mechanical and biological properties. Contribution was proving that ML can replace costly simulations while still generating clinically relevant, customizable bone scaffold designs.
Conev et al., 2020	Introduced machine-learning-guided 3D printing for tissue engineering scaffolds. Algorithms were trained on material type and printing parameters to predict scaffold quality metrics, including porosity, printability, and fidelity. Demonstrated improved reproducibility and reduced trial-and-error during scaffold fabrication.	Key contribution: Providing a data-driven fabrication model that directly reduces experimental workload and enhances fabrication consistency, demonstrating that ML can control scaffold printing workflows in real time.
Mackay et al., 2021	Explored AI integration in bone regeneration and TE, demonstrating examples like conditional GANs for image synthesis, and random forest models predicting dissolution rates of bioactive glass (r ≈ 0.99). Also described is the computational analysis of angiogenesis via the CAM assay. Highlighted the importance of AI in linking materials science datasets with biological outcomes,	Provided detailed case studies of how AI models optimize material degradation, angiogenesis quantification, and imaging analysis in TE. The contribution was presenting AI as a bridge between materials informatics and biological assessment, laying the groundwork for predictive TE.
Malekpour & Chen, 2022	Comprehensive review on printability and cell viability in extrusion bioprinting. Highlighted how Gaussian process regression, anomaly detection algorithms, and multi-objective optimization approaches are applied to predict outcomes like cell survival, filament diameter, and construct uniformity, and collected data across multiple studies to illustrate algorithm performance.	Key contribution was cataloguing and standardizing the ML techniques used in bioprinting optimization, offering reproducible frameworks for future TE research. Emphasized that ML can effectively reduce bioprinting variability and improve the reproducibility of constructs.
Ding et al., 2024	Review on quantum dots (QDs) for bone tissue engineering, describing their optical properties, biocompatibility issues, and applications in imaging, sensing, and scaffold functionalization. Discussed QDs as potential components for theranostic scaffolds, enabling real-time monitoring of bone regeneration.	The contribution introduced quantum dot-enabled TE as a novel hybrid approach, in which QDs serve both diagnostic (imaging) and therapeutic (drug delivery/monitoring) roles within scaffolds.
Vaishya & Haleem, 2022	Brief commentary on technology adoption by orthopaedic surgeons, with reflections on how emerging AI and digital systems are shaping surgical workflows. Although not TE-specific, it emphasizes the hurdles clinicians face in accepting and adopting.	The contribution was to contextualize TE and regenerative AI innovations within the practical realities of clinical uptake, stressing that without clinician buy-in, translational success is limited.
Logeshwaran et al., 2024	The review demonstrated that AI-assisted 3D printing optimizes scaffold porosity, geometry, and printability for patient-specific bone regeneration. CNNs and ResNet models achieved ∼92.1% accuracy in detecting defects such as fractures and porosity, while ANNs predicted scaffold mechanics and porosity under different printing parameters. Optimization tools, including PSO and pareto analysis, identified ideal printing settings, and the combination of open-loop and closed-loop AI enabled predictive corrections and real-time quality control.	The study emphasized that AI integration reduces errors, costs, and processing time while producing scaffolds with tunable mechanical and biological properties. It highlighted AI’s potential to replace animal models in preclinical testing and positioned AI-driven scaffold fabrication as a pathway to reproducible, vascularized, and personalized bone constructs.

Other research described how AI can be integrated into scaffold design, CAD modelling, process monitoring, and biological evaluation.^22,25,28 AI can also accelerate both the design and evaluation phases. A study by Fredrik Liu et al. quantified MSC dosage and its effects in depth, offering predictive insights for cell-based therapies. ²⁹

AI also extends into intelligent manufacturing systems. Lyu et al. proposed AI-assisted production of osteoarthritis organoids, combining automation, predictive modelling, and quality control. ³⁰ This framework aims to stabilize OA microenvironments in vitro, producing reproducible organoids that support drug testing and disease modeling ³⁰

More recently, Logeshwaran et al. reviewed artificial intelligence-based 3D printing strategies for bone scaffold fabrication, demonstrating how AI can enhance scaffold porosity, topology, geometry, and printability to produce patient-specific constructs. ³¹ The study highlighted the use of convolutional neural networks (CNNs) and ResNet models for quality control, with the capability of detecting fractures, gas porosity, and lack of fusion with ∼92.1% accuracy. In contrast, artificial neural networks (ANNs) were employed to predict scaffold mechanical properties and porosity under varying printing parameters. Optimization methods, such as particle swarm optimization (PSO) and Pareto front analysis, were employed to identify ideal printing conditions. The combination of open-loop and closed-loop AI approaches enabled both pre-printing geometry programming and real-time corrective feedback. Importantly, the review emphasized the potential of AI-assisted 3D scaffolds to serve as alternatives to animal models in preclinical studies, thereby reducing ethical concerns while supporting translational applications. ³¹

Personalization of cell-based therapies

AI-driven network analysis has become a powerful tool for narrowing down candidate genes and pathways from high-throughput datasets. For example, Shi et al. integrated transcriptomic profiling with protein-protein interaction (PPI) networks in rheumatoid arthritis and identified hub genes such as LILRB1 and ATXN1, which subsequently guided structure-based drug screening. ³² In a similar approach, Huo et al. combined ferroptosis-related networks with differentially expressed genes (DEGs) in osteoporosis, uncovering 10 hub genes central to bone metabolism and iron handling. ³³ These computational pipelines effectively condense thousands of genes into a focused set of candidates that can be directly leveraged for drug discovery and docking workflows.

Personalization has also emerged as a unifying theme across recent studies. Liu et al. demonstrated that AI can optimize mesenchymal stem cell (MSC) dosing, identifying optimal dose ranges, critical defect size thresholds, and case-specific confidence measures. ²⁹ Such predictive models reduce empirical trial-and-error and enable patient-tailored regenerative medicine. More broadly, reviews emphasize the importance of embedding personalization into imaging, risk stratification, and therapy planning in orthopaedics. Underscoring AI’s role in advancing precision medicine^21,23,24

Drug discovery and delivery

In orthopaedics, AI has helped accelerate drug discovery by integrating omics-derived targets with cheminformatics data, docking, and materials-enabled delivery platforms (Table 7). In rheumatoid arthritis, a study done by Shi et al. screened nine compounds, including luteolin, naringenin, apigenin, and quercetin-3-glucoside, against hub targets such as LILRB1 and DLG2. Binding affinities below −6.0 kcal/mol, supported by stable hydrogen bonding and π–π interactions, highlighted these as promising candidates. ³² In osteoporosis, Huo et al. applied docking and pharmacology analysis to prioritize quercetin, resveratrol, epigallocatechin, and other natural compounds, showing high binding affinity to ferroptosis-related targets. ³³

Table 7.

Drug discovery and delivery in orthopaedic research.

Reference	Results	Key contributions
Shi et al., 2020	Used bioinformatics analysis of transcriptomes from rheumatoid arthritis patients to identify differentially expressed genes (DEGs) and construct PPI networks. Found LILRB1 as a hub gene. Conducted drug screening via molecular docking, identifying candidate compounds such as kaempferol glucoside with favorable binding affinities.	Contribution was developing a systematic in silico pipeline for drug discovery in RA: (1) identify DEGs, (2) build PPI networks, (3) select hub genes, (4) perform molecular docking. Demonstrated how AI-supported pipelines can identify novel therapeutic targets and candidate molecules in autoimmune joint disease.
Huo et al., 2024	Performed integrated network pharmacology and docking for ferroptosis-related osteoporosis. Identified five hub genes (TP53, EGFR, TGFB1, SOX2, MAPK14) via PPI analysis. Enrichment analysis (GO, KEGG) linked them to oxidative stress pathways. Docking (with AutoDock vina, AlphaFold models) prioritized compounds such as resveratrol, quercetin, apigenin, and hydroxytyrosol as promising multi-target agents.	The contribution was the presentation of a complete computational discovery pipeline that links ferroptosis pathways to osteoporosis progression and provides polyphenol candidates for therapeutic intervention. Demonstrated how AI + docking can reveal multi-target, multi-pathway treatments in bone loss disorders.
Yang et al., 2020	Developed probabilistic neural network models to predict outcomes of MSC therapy for cartilage repair. Input features included lesion size, depth, species, MSC dose, and injury type. Achieved performance of R2 = 0.637 ± 0.005, and identified size/depth thresholds predictive of treatment efficacy.	The contribution was the development of a quantitative predictive tool for MSC therapy outcomes. Provided decision thresholds for when MSC therapy is most likely to succeed, guiding patient selection and treatment planning in cartilage regenerative medicine.
Lei et al., 2023	Reviewed materials-based therapeutic strategies for osteoporosis, emphasizing scaffold-mediated drug delivery, nanoparticles, hydrogels, and bioactive glasses. Discussed innovations such as ROS-scavenging nanomaterials, acid-neutralizing agents, and immunomodulatory biomaterials. Outlined challenges such as burst release and local inflammatory responses.	The contribution was the compilation of state-of-the-art drug delivery platforms for bone regeneration. Highlighted how biomaterial carriers improve therapeutic retention and bioactivity compared to systemic drug administration.
Ding et al., 2024	Summarized quantum dots in bone TE, particularly their use as drug carriers and bioimaging agents. Highlighted how QDs can be integrated into drug-loaded scaffolds, enabling theranostic scaffolds with simultaneous imaging and delivery functions.	The contribution described multifunctional quantum dots as dual-purpose tools for regenerative drug delivery, capable of tracking delivery and enhancing therapeutic control in osteogenic scaffolds.

Complementary to molecular docking, materials-based platforms extend drug discovery into the realm of delivery. Based on the research of Lei et al., reviewed biomaterials, such as hydrogels, nanocarriers, and scaffolds, that help in more targeted and sustained delivery of osteoanabolic and osteoimmunomodulatory agents. ³⁴ Whereas research done by Ding et al. described Quantum Dots (QDs) as a theragnostic tool, that enables or allows the combination of drug delivery with imaging and osteogenesis-enhancing functions. Although it was mentioned that biocompatibility and toxicity remain critical challenges that need further research and testing, ³⁵ This continuum highlights how AI connects drug discovery with translational strategies in orthopaedics.

Genomics and precision orthopaedics

While advancements in physical tools have become popular, data and science provide the intelligence to guide their use. In this case, genomics offers a foundational role (Table 8). Integrating genomic science into orthopaedic care has enabled precise, targeted care. These advancements improved the care for commonly concerning orthopaedic diseases, such as osteoarthritis (OA), osteoporosis, and scoliosis. For instance, Genome-Wide Association Studies (GWAS) have revealed associations between genetic polymorphisms and OA. Aubourg et al. ³⁶ identified over 100 polymorphic DNA variants associated with OA, accounting for more than 20% of its heritable component. These findings underscore the importance of understanding the genetic contributors to OA and highlight the potential of targeted therapies to prevent the genetic risks. ³⁶

Table 8.

Genomics and precision orthopaedics.

Reference	Focus area	Key contributions
Li G et al., 2024	Proteomics in orthopaedic research and translational implications.	Proteomics holds considerable promise for identifying biomarkers to detect and treat orthopaedic diseases early.
Wang T et al., 2023	Single-cell RNA sequencing in orthopaedic research.	Single-cell RNA sequencing provides information about the cell types that make up bone and cartilage, helping researchers develop personalized treatments.
McKinley TO et al., 2019	Precision medicine for orthopaedic trauma management.	Precision medicine can customize treatments for trauma patients, enhancing recovery according to genetic predisposition.
Luo Y et al., 2024	Fully automated personalized orthopaedic treatments.	Automating orthopaedic treatments, with a focus on individualized care, has the potential to address existing deficiencies in treatment efficiency and accessibility.
Miao J et al., 2024	Machine learning-assisted GWAS in orthopaedic genetics.	Machine-learning-assisted GWAS could identify genetic variants that affect orthopaedic conditions, enabling targeted genetic interventions for each individual.
Tanikella AS et al., 2020	Gene-editing for osteoarthritis.	Gene editing could help with osteoarthritis by enabling changes to joint tissues that are unique to each person.
Choi Y et al., 2019	Genome engineering for osteoarthritis.	Genome engineering methods for osteoarthritis can help develop drugs and stem cell-based treatments tailored to each person.
Kang YJ et al., 2021	Differential gene expression in osteoporotic hip fracture patients with sarcopenia.	Gene expression variations in elderly osteoporotic patients indicate the necessity for tailored therapies to address fractures in aging demographics.

Aside from genomics, proteomics offers real-time insights into disease by revealing the dynamic functional states of tissues. By using new biomarker panels and identifying disrupted pathways across various biological samples, proteomics complements structural data derived from both genetic and imaging analyses. It enhances understanding of disease mechanisms, stratifies patient risk, monitors recovery, and predicts treatment response. ³⁷ The omics can be discovered from single-cell RNA sequencing (scRNA-seq), a powerful tool for investigating cellular diversity and mechanisms underlying orthopaedic diseases. This advancement could lead to the development of novel therapeutic targets, supporting precision orthopaedic practices. ³⁸

The PRECISE STUDY by McKinley et al., ³⁹ is a leading example of comprehensive data integration. Their “precision injury signature” concept, developed innovatively, created a multidimensional profile for patients with multiple injuries by analysing data on mechanical damage, blood flow dynamics, immunological responses, and demographics. Furthermore, understanding molecular interactions at the genomic level facilitates the design of effective bioactive scaffolds, particularly in the construction of implants and joint reconstruction. ⁴⁰

Artificial intelligence enables more effective analysis of GWAS data, offering new avenues for uncovering genetic insights. These insights enhanced the identification of genetic risks and facilitated the prediction of susceptible diseases based on phenotypic expression. ⁴¹ Additionally, these genetic insights are also helping the development of targeted therapies. For instance, gene-editing techniques enable direct modification of genetic sequences associated with OA pathology.^42,43 Metagenomic sequencing can also analyse the pathogens of bone and joint infections. A study by Maimati et al. ⁴⁰ demonstrated 16S rRNA sequencing has a high concordance rate (87.5%) with bacterial cultures in detecting pathogens. This approach also recognised overlapping genetic pathways, for example, sarcopenia and osteoporosis both contribute to an increased risk of disability. ⁴⁴

AI and prosthetic limb control

Lower limb prosthetics and exoskeletons have evolved from passive mechanical devices to powered, microprocessor-controlled systems. ⁴⁵ Despite technological advances, intuitive and reliable user-device interaction remains a significant challenge, particularly in replicating natural volitional control.^46,47 AI plays a vital role in replicating natural volitional control in prostheses and exoskeletons (Table 8).

Advanced machine learning techniques can analyze electromyography (EMG) signals and intuitively control the prosthetic device, thereby reducing the user’s cognitive burden. In a case series by Edwards et al., the application of General Value Functions (GVF) can predict a user’s intended next move in myoelectric upper-limb prostheses. The study demonstrated that predicting the user’s intended next move reduces the time and cognitive load required for amputees to operate their prosthetic device. ⁴⁵ For lower limb prosthesis or exoskeleton, Cimolato et al. and Belal et al. conduct a systematic review evaluating machine learning processing to enhance the lower limb exoskeleton and prosthesis control.^46,47 Cimolato et al. found that pattern recognition is the most prevalent approach for microprocessor-controlled lower limb prostheses (MLLPs). ⁴⁶ The key findings from a study by Belal et al. on deep learning algorithms for lower limb exoskeletons included gait phase recognition, locomotion mode prediction, and torque/angle estimation. ⁴⁷ CNNs, LSTM models, and reinforcement learning were most widely applied, demonstrating superior accuracy compared to traditional rule-based approaches. ⁴⁷

From this literature, EMG remains essential for capturing user intent. However, hybrid systems integrating predictive algorithms and multimodal sensors hold tremendous promise for reliable, intuitive, and adaptive control.^45–47

Synthetic data and virtual trials

Advances in technology, particularly AI and robotics, have significantly expanded the scope of medical innovation in recent years. A major challenge in research, however, remains data scarcity – especially in rare conditions or studies requiring strict compliance.^13,48 One emerging solution is the use of synthetic data, generated through deep or machine learning, to augment or substitute for real datasets. ⁴⁹ In orthopaedics, multiple studies have explored the generation of synthetic data for medical imaging, motion analysis, and biomechanics applications (Table 9).

Table 9.

AI and prosthetic limb control.

Reference	Method/Device	AI method	Main outcomes	Notes
AI in prosthetic
Cimolato et al. (2022)	Systematic review of EMG-driven controllers in microprocessor-controlled lower limb prostheses (MLLPs); classified by CIC vs IEC, and by control principles (direct, pattern recognition, model-based).	Pattern recognition, model-based control, direct EMG control.	EMG enables volitional and rhythmic locomotion control; pattern recognition is most frequently applied; hybrid control strategies (combining multiple methods) may improve performance; highlighted importance of volitional intent decoding.	No standardized performance metrics; absence of validated clinical translation; commercial products not yet available.
Edwards et al. (2015)	Case series (1 amputee, 4 non-disabled); robotic arm controlled under adaptive vs conventional switching conditions.	Real-time prediction learning using General value functions (GVFs) (Reinforcement learning paradigm).	Adaptive switching reduced the number of switches and switching time; decreased cognitive load for the user; demonstrated the feasibility of real-time ML to anticipate user intent.	Small sample size; focus on upper limb prosthesis, limiting direct generalizability.
Irshaid et al. (2024)	Review of 103 studies (2018–2023); covered gait recognition, prediction, torque/angle estimation.	Deep neural networks (DNNs), including CNNs, RNNs (LSTM), transformer-based models, and deep reinforcement learning.	DL approaches outperform traditional methods for gait phase detection, locomotion intent recognition, and joint torque/angle estimation; multimodal sensor fusion (EMG, IMU, force sensors, vision) improves robustness.	Lack of standardization in evaluation metrics; most studies in lab settings; limited clinical validation.

Gao et al. introduced SyntheX, a framework for generating realistic simulated images from human models. ⁴⁹ This system enables the development of generalizable AI algorithms for X-ray image analysis. Their study compared models trained on realistically simulated data (Sim2Real) with those trained on SyntheX-generated synthetic data and found comparable performance. These results suggest that synthetic data from SyntheX can serve as a viable alternative—or complement—to real data acquisition. ⁴⁹

Motion analysis and biomechanics are essential for comparing healthy and pathological individuals and assessing disease progression. However, human motion capture research is challenging because it requires considerable time commitments and significant financial resources. Thus, with the development of deep learning, generative AI can be employed to augment datasets by generating various synthetic biomechanical data. Perrone et al. (2025) applied a variational autoencoder with temporal modelling to generate synthetic kinematic and kinetic variables for motion analysis. ⁴⁸ Both real data alone and a combination of real and synthetic data were used to train LSTM models. The performance of LSTM models trained on real data alone and on a combination of real and synthetic data is comparable, with normalised root mean square errors (nRMSE) on all four biomechanical outputs. ⁴⁸

Sharifi Renani et al. (2021) further advanced this field by generating synthetic inertial measurement units (IMUs) to train recurrent neural networks for predicting 3D joint angles during gait. ¹³ Using a musculoskeletal modeling workflow in OpenSim, they augmented real IMU data with synthetic signals. Their results showed substantial improvements: prediction accuracy increased by 38% at the hip and 11% at the knee, while RMSE was reduced by 54% and 45% at the hip and knee, respectively. These findings highlight the potential of synthetic data to enhance predictive modeling where real data are limited. ¹³

Taken together, these studies demonstrate that synthetic data generation offers a powerful strategy to address data scarcity in orthopaedic research. Whether in imaging, motion analysis, or biomechanics, synthetic datasets can improve model robustness, reduce reliance on costly real-world data collection, and accelerate the development of AI-driven medical technologies. However, in data-scarce settings, synthetic data generation may lack the complexity of real-world variability and omit necessary factors.

Translational AI for orthopaedic diagnostics

Machine learning and deep learning

Machine Learning and Deep Learning are computer techniques that learn patterns from orthopaedic data like X-rays, CT/MRI scans, or clinical records to help clinicians spot problems sooner, measure them more consistently, and predict risks that matter for treatment and recovery. ⁵³

Machine Learning uses algorithms that look at input features (for example, age, bone measurements, or image-based markers) to classify or predict outcomes.^54,55 Deep Learning is a specialized type of Machine Learning that uses multi-layer neural networks to automatically learn patterns directly from raw images or signals, which is especially helpful for tasks such as detecting fractures, grading osteoarthritis, segmenting bones and joints, and recognizing implants or soft-tissue injury in complex 2D–3D imaging.^53,56 In everyday practice, ML/DL can speed up image reading by auto-segmenting the spine or hip, flag likely bone metastases on CT, or screen for low bone density supporting faster decisions and more personalized care alongside clinician judgment. ⁵⁷ Modern reporting checklist such as Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD + AI) for prediction models and Checklist for Artificial Intelligence in Medical Imaging (CLAIM) for medical-imaging AI explain what authors must disclose about data, model training, testing, and limitations so clinicians can trust what they read and reproduce it. Beyond headline accuracy, models should also report calibration because these determine safe decision thresholds and triage rules in clinical workflows.^37,57

Recent studies by Hess et al. and Xiao et al. have highlighted the growing role of deep learning in advancing these imaging techniques.^50,51 By leveraging complex pattern-recognition and predictive capabilities, machine learning provides powerful tools to enhance diagnostic accuracy, improve efficiency, and address challenging clinical cases in orthopaedics (Table 10).

Table 10.

Synthetic data and Virtual trials in orthopaedic research.

Reference	AI methods	Key findings	Notes & limitations
Gao et al. (2023)	Simulation from CT (SyntheX), domain randomization, domain adaptation (CycleGAN, ADDA).	Models trained on synthetic radiographs achieved performance comparable to or superior to real-data models in hip imaging, robotic tool detection, and COVID-19 lesion segmentation; scaling synthetic data further improved performance.	Domain gap persists; realism of simulated data is critical; requires domain transfer methods to ensure robustness.
Perrone et al. (2025)	Variational autoencoder (VAE) with LSTM.	Synthetic kinetic & kinematic variables statistically indistinguishable from real data; LSTM models trained on real + synthetic data showed lower nRMSE vs real-only models (e.g., knee sagittal moment: 9.21% vs 9.75%).	Small sample size (n = 53); synthetic data may not capture complex inter-subject variability; validation limited to single-leg squat task.
Sharifi renani et al. (2021)	Musculoskeletal modeling + neural networks (recurrent).	Synthetic IMU signals improved prediction accuracy: Hip RMSE reduced from 4.5° (real-only) to 1.9° (synthetic + real); knee RMSE from 3.3° to 1.7°; models trained on synthetic-only data performed well.	Synthetic IMUs may not replicate soft-tissue noise; limited to gait tasks; real-world generalizability not fully validated.

Lumbar spine MRI

Hess et al. successfully applied machine learning to perform fully automated quantitative imaging from lumbar spine MRI data. In their study, convolutional neural networks (V-Net–style) were trained to segment vertebral bodies, intervertebral discs, and paraspinal muscles on routine sagittal and axial T1-weighted MRI scans. The AI-generated masks showed high agreement with human annotations, achieving segmentation Dice coefficients ≥0.77, and strong correlations for derived morphometric and biomechanical parameters (≥0.69 when significant). These findings suggest that AI-based segmentation could serve as a reliable substitute for manual annotations of major lumbar spine structures, with potential for integration into future clinical workflows ⁵⁰

CT-scan for femoral neck anteversion (FNA)

Until recently, the assessment of femoral neck anteversion (FNA) has been limited by substantial intra- and inter-observer variability. Xiao et al. addressed this challenge by developing a fully automated FNA workflow based on deep learning. ⁵¹ Their model integrates femur segmentation, landmark identification, and anteversion calculation. Two-axis definitions were employed: Auto_GT, using the greater trochanter relative to the intercondylar notch center, and Auto_P, using the piriformis fossa relative to the intercondylar notch center. The deep learning approach achieved high measurement consistency, with intraobserver intraclass correlation coefficients (ICCs) ranging from 0.864 to 0.961. These results demonstrate that an automated, 3D CT–based method can provide reliable and highly reproducible FNA measurements, offering a robust alternative to conventional manual techniques.

Knee OA detection from MRI

Knee osteoarthritis (OA) is inherently a three-dimensional disease process, requiring volumetric imaging for early and accurate detection. Yeoh et al. addressed this challenge by adapting 2D convolutional neural network (CNN) weights to 3D, creating transfer-learning–assisted 3D CNN architectures (ResNet, DenseNet, VGG, AlexNet) trained on MRI data from the Osteoarthritis Initiative. ⁵² Across 13 tested architectures, transfer learning markedly enhanced performance, particularly in ResNet and DenseNet models. Among these, ResNet34 achieved the highest accuracy (0.875) and F1 score (0.871), while shallower networks such as ResNet18 and DenseNet121 reached AUCs of 0.945 and 0.914, respectively. These findings highlight the ability of 3D CNNs to capture the complex anatomical and structural features of the knee joint, supporting their potential use as computer-aided diagnostic tools for OA when volumetric MRI data are available. ⁵²

Automatic detection and classification of bone metastases

Bone metastasis (BM) detection on CT scans is a time-consuming process, traditionally focused primarily on the spine. To address this limitation, a multi-institutional study developed an automated Bone Lesion Detection System (BLDS) trained on 2,518 CT scans encompassing 9,177 BMs from five hospitals. ⁵³ The BLDS achieved 89.1% lesion-wise sensitivity at 1.40 false positives per case and delivered 92.3%/91.1% accuracy for BM versus non-BM classification in internal and external test sets, respectively. Integration with radiologist workflows improved lesion-wise sensitivity by 22.2% and reduced reading time by 26.4%. Real-world validation in 54,610 patients further confirmed strong performance, with 90.2% patient-level sensitivity and a 98.2% negative predictive value, underscoring its readiness for clinical deployment. ⁵³

Opportunistic osteoporosis detection using LDCT

A 2025 study developed an opportunistic screening model for osteoporosis using low-dose abdominal CT. ⁵⁴ he approach combined deep learning–based automatic segmentation of the proximal femur (VB-Net) with radiomics-based classification. Segmentation performance was excellent, with Dice similarity coefficients of 0.975 ± 0.012 (validation) and 0.955 ± 0.137 (test). For bone status classification, a random forest radiomics model achieved strong discrimination: AUCs of 0.924 for normal bone mass, 0.960 for osteoporosis, and 0.828 for osteopenia. Corresponding sensitivity and specificity values were high, with osteoporosis detection reaching 0.947 and 0.963, respectively. This framework demonstrates the potential of leveraging routine CT scans for radiation-free, cost-neutral, and opportunistic AI-based osteoporosis screening. ⁵⁴

Hip fracture risk prediction with uncertain quantification

Shaik et al. developed a predictive model for hip fracture risk by integrating clinical variables with imaging features derived from hip dual-energy X-ray absorptiometry (DXA). CNN-extracted features included shape and texture measurements, which were incorporated into two ensemble models: Ensemble 1 (clinical variables only) and Ensemble 2 (clinical plus imaging features). Ensemble 2 demonstrated superior performance, achieving an AUC of 0.95, accuracy of 0.92, sensitivity of 0.81, and specificity of 0.94. Furthermore, a staged workflow with uncertainty quantification enabled approximately 54% of patients to be ruled out from requiring DXA scans, while maintaining high diagnostic accuracy. This approach highlights a cost-effective and patient-specific strategy for fracture risk prediction and screening optimization. ⁵⁵

Common challenges, emerging trends, and consensus across studies

Across the studies included in this review, several recurring challenges and limitations emerge in the current application of AI in orthopaedic research. These include a limited dataset, a lack of standardized protocol, insufficient external validation, and substantial heterogeneity in evaluation metrics. Data scarcity and limited data variables affect research areas such as gait analysis, micro-CT imaging, regenerative modelling, and prosthetic limb control.^12–15 Due to scarcity and limited data variables, many machine learning or deep learning models are trained on single-center or simulation-based datasets, leading to limited generalizability and increased risk of overfitting.^12–14 Several studies also highlight the lack of a standardized protocol for AI development and evaluation in orthopaedic research. For instance, multiple gait analysis studies highlight a systemic lack of harmonized protocols for hardware configuration, sensor placement, gait task instructors, video acquisition settings, annotation methods, and preprocessing pipelines.^7,10,11,13 Another major concern is the insufficient level of external validation in most imaging and biomechanical studies. Some AI models included in this review have not been deployed in the real world. Hence, limiting their translational ability. Moreover, reporting performance metrics such as AUC, RMSE, or Dice coefficient remains inconsistent across studies. Thus, underlining the importance of unified reporting frameworks such as TRIPOD-AI and CLAIM.^56,57 Finally, interpretability represents a persistent challenge in deep learning models. The use of various algorithms with limited transparency may hamper clinical trust and regulatory approval; this issue is highlighted by the studies of Kolomenskaya et al. and Farhadi et al.^22,23

Despite the challenges identified in AI orthopaedic research, several notable trends have emerged, including the use of synthetic data and generative modelling, a shift toward multimodal and integrative approaches, advancements in regenerative orthopaedics, and the development of real-time intelligent systems. Synthetic data and generative modelling have shown promising results, as demonstrated by synthetic IMU generation and syntheX.^13,49 Both studies demonstrated that simulated data can meaningfully improve and reduce the prediction error of an AI model. By combining video, IMU, EMG results, radiomics, genomics, and proteomics, diagnostic precision and decision support in orthopaedics can be enhanced.^8,11,36,37 In the field of regenerative orthopaedics, AI can aid in optimizing scaffold design, generate predictive models for mesenchymal stem cell therapy, and automate organoid fabrication. ²⁴ Through predictive modelling, AI can also reinforce prosthetic and exoskeleton control. Machine learning, such as LSTM and reinforcement learning, helps translate user intent into more intuitive device behaviour.^45,46 Collectively, AI in orthopaedic research does not replace conventional orthopaedic methods. Rather, it functions as a powerful tool that enhances accuracy, accelerates workflows, and promotes more integrated and individualized care. However, broader multicentre studies, consensus on regulatory framework, and integration strategies for safe and effective deployment must be implemented before clinical translation can be fully realized.

Research gaps and future research directions in AI-driven orthopaedics researches

Despite rapid advances, significant research gaps persist that limit the scalability and translational readiness of AI applications in orthopaedics research. Most notably, the lack of large, diverse, and multi-institutional datasets poses a major obstacle across imaging, gait biomechanics, prosthetic control, and tissue engineering.^12–15 Many existing models are developed using controlled laboratory or simulation environments, resulting in limited external validity and uncertain real-world performance. The majority of studies also lack rigorous external validation, with only a small proportion evaluating models across independent populations or clinical settings, as highlighted by research on bone metastasis detection and opportunistic osteoporosis screening.^53,54 In regenerative orthopaedics, translational pipelines are still fragmented, with limited continuity between in silico scaffold design, in vitro organoid modelling, in vivo validation, and clinical readiness. ²⁴

For further advancement of AI in orthopaedic research, future studies should develop an organized multicenter data networks that facilitate collaborations while increasing the diversity and representativeness of training datasets. Harmonizing evaluation frameworks through systematic adoption of TRIPOD-AI, CLAIM, and ISO/IEC 23894 will be essential to ensure transparent reporting, robust calibration, and reliable external testing.^56,57 Future research should also invest in trustworthy and interpretable AI, enabling clinicians to understand the AI model’s behaviour across imaging, biomechanics, and regenerative medicine, thereby supporting clinical acceptance and regulatory approval. Integrating multimodal data, including genomics and proteomics with gait, imaging, and biomechanics evaluation, may further enhance precision in orthopaedic research.^8,11,36,37 Combining finite element modelling, multimodal sensor data, and patient profiles could enable more personalized implants and more accurate fracture-healing prediction.^6,9,10 In regenerative medicine, progress will benefit from AI-driven scaffold design, automated organoid fabrication, and standardized preclinical validation frameworks.^21–24,29 Finally, regulatory-ready AI will require ongoing research into fairness auditing, robustness testing under domain shift, lifecycle monitoring, and real-world post-deployment evaluation. Together, these directions outline a roadmap toward more reliable, explainable, and clinically impactful AI systems in orthopaedics.

Comparison of artificial intelligence models in orthopaedic research

Machine Learning (ML) and Deep Learning (DL) can serve as complementary or assist in orthopaedic research. Current ML algorithms, such as Decision Trees, Random Forests, Support Vector Machines, and Gradient Boosting, are considered important for certain research where datasets are structured or limited. These algorithms or methods are prized for transparency, lower computational demands, and robustness in situations with limited data.^58,59

They require relatively low computational resources and provide transparent decision-making, which is valuable for applications such as biofabrication parameter optimization and clinical risk modelling. Conev et al. used Random Forest classifiers to predict print quality in 3D-printed scaffolds, enabling precise differentiation of optimal printing parameters and minimizing potential redundancies in experimentation. ²⁶ Similarly, Malekpour and Chen combined experimental and ML algorithms to help in the assessment of printability and cell viability in extrusion-based bioprinting, which demonstrates the effectiveness of ML-assisted control in tissue engineering. ²⁷ These studies highlight ML’s strengths in interpretable, parameter-driven, and data-limited applications.

Deep Learning (DL), which includes Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and transformer-based models, can automatically extract hierarchical features from unstructured data, enabling end-to-end learning from complex imaging or histological data without manual feature engineering. These DL approaches have achieved diagnostic accuracies ranging from 85% to 99% in orthopaedic applications, including fracture detection, implant recognition, and osteoarthritis grading. ⁶⁰ Gao et al. developed the SyntheX framework, which trained DL models on synthetic X-ray datasets and reported equivalent or superior performance when compared to models trained on real clinical data for hip imaging and surgical tool detection. ⁴⁹ Likewise, Paek et al. integrated DL-assisted image analysis into a bone-on-a-chip platform to evaluate the efficacy of an anti-SOST antibody. They reported achieving over 95% accuracy in β-catenin translocation patterns. ¹⁷ However, these performance advantages are not without disadvantages. It is reported that there are challenges when using said algorithms, which have higher computational requirements and limited interpretability, as DL models often function as “black boxes,” complicating their clinical translation and regulatory validation.^17,60

Supervised learning, which relies on labeled datasets, remains the dominant paradigm in both ML and DL applications, consistently achieving high predictive accuracy in classification and regression tasks.^17,49,59,61 It has been especially successful in orthopaedic applications involving image segmentation, motion tracking, and cellular-level imaging. By contrast, unsupervised learning discovers hidden structures and correlations in unlabeled data via clustering or dimensionality reduction (e.g., K-Means, PCA). Less utilized in orthopaedics due to a lack of standardized evaluation metrics, related domain studies suggest that unsupervised or self-supervised pretraining can achieve 15-30% performance improvements in data-limited conditions. ⁶¹ This suggests applicability for orthopaedic datasets where annotations may be rare or inconsistent.

While small datasets are a potential limitation to any machine learning technique, TL can practically adapt pre-trained models or synthetic data to orthopaedic imaging tasks.^49,60 Medical imaging using TL has demonstrated near-perfect accuracy while reducing the need for data and computation. ⁶⁰ Gao et al. have shown that TL in a synthetic data generation framework, called SyntheX, enables deep networks to generalize well from simulated to real radiographic data with accuracy comparable to that of models trained solely on real data. ⁴⁹ By contrast, training from scratch offers a lot of flexibility in architecture but remains impractical in orthopaedics because of limited data availability and high computational costs.^17,49,60

In summary, ML models are superior in terms of interpretability and efficiency with smaller, structured datasets, while DL models thrive with large, complex datasets. Supervised algorithms are known to achieve high accuracy in well-annotated or labelled data, whereas unsupervised or hybrid supervised-unsupervised algorithms offer flexibility for unlabelled data. TL algorithm model strikes a practical balance by improving accuracy and generalization in low-data contexts. These models (ML, DL, and TL) each have their own advantage and disadvantages, so future researcher must have have a clear idea of what type of data that is being used or interpreted and use the appropriate algorithm that suits the research that is being conducted and can not be blindly used to enhance accuracy, consistency, and translational results in orthopaedic studies.

Ethical, privacy, and regulatory considerations in translational orthopaedic AI

Despite rapid progress, translating AI into orthopaedic research and clinical workflows requires careful attention to data governance, patient privacy, algorithmic transparency, and regulatory compliance. Preclinical and clinical datasets often contain identifiable imaging, gait signatures, and genomic patterns, making privacy-preserving data handling essential. Regulatory frameworks such as HIPAA (U.S.) and GDPR (European Union) require secure processing, de-identification, and controlled sharing of patient data, while emerging global frameworks increasingly emphasize data minimization and traceability of model decision-making.

For medical AI software, multiple governing standards outline validation, risk management, and reporting expectations. In particular, the FDA’s Software as a Medical Device (SaMD) framework and the EU Artificial Intelligence Act (2024) classify clinical decision-support models as high-risk systems requiring external validation, post-deployment monitoring, and bias auditing.^62,63 Professional organizations have published complementary methodological standards, including TRIPOD + AI for transparent reporting of prediction models and CLAIM for AI in medical imaging, both of which promote reproducibility, calibration reporting, and external testing.^56,57 Additionally, ISO/IEC 23894:2023 and IEEE 7001/7003 provide governance principles for algorithmic fairness and human oversight. ⁶⁴ Hence, multicentric studies and real-world data validation are needed to ensure the AI systems can be reliably translated into clinical settings.

Recent studies increasingly adopt privacy-preserving computation approaches, including federated learning (training across multiple centers without centralizing data) and differential privacy to mitigate re-identification risks. ⁶⁵ Synthetic data generation and domain-randomized simulation platforms further support model training where real patient data are limited, while also reducing exposure of sensitive datasets. However, challenges remain, particularly regarding dataset representativeness, explainability of complex deep learning architectures, and regulatory harmonization across regions. Addressing these concerns will be essential for enabling safe and equitable deployment of AI-driven orthopaedic technologies from laboratory settings to clinical practice (Table 11).

Table 11.

Translational imaging.

Reference	AI methodology	Key findings	Notes & limitations
Du et al. (2025)	VB-net deep learning for proximal femur segmentation + radiomics with random forest classifier.	Achieved DSC 0.975 (validation) and 0.955 (test) for segmentation; radiomics model AUC 0.960 for osteoporosis classification; high sensitivity/specificity across normal, osteopenia, and osteoporosis groups.	Automated LDCT-based segmentation and radiomics can provide cost-effective opportunistic screening for osteoporosis.
Shaik et al. (2024)	Ensemble ML models with logistic regression, CNN-extracted imaging features, and uncertainty quantification (staged model).	Ensemble 2 (clinical + DXA features) achieved AUC 0.95 and accuracy 0.92; the staged model reduced DXA need by ∼54% while maintaining AUC 0.85.	A staged ML approach can accurately predict hip fracture risk while reducing unnecessary DXA scans, lowering cost and radiation exposure.
Yeoh et al. (2023)	Transfer learning–assisted 3D CNNs (ResNet, DenseNet, VGG, AlexNet) applied to MRI.	ResNet34 best (accuracy 0.875, F1 0.871); shallow networks (ResNet18, VGG11) yielded high AUC (0.945, 0.928). Transfer learning improved performance with limited MRI data.	3D CNNs with transfer learning are promising for knee OA diagnosis and are recommended for clinical aid, even with limited datasets.
Zhang et al. (2025)	Bone lesion detection system (BLDS) using deep CNNs trained on multicenter CT datasets.	BLDS sensitivity 89.9% (1.4 FP/case); BM/non-BM classification accuracy ∼92%. Outperformed radiologists in lesion detection and reduced reading time by 26.4%.	BLDS can augment radiologists, improving BM detection accuracy and efficiency in routine CT workflows.
Hess et al. (2023)	CNN-based segmentation (2D V-nets) on lumbar spine MRI for vertebrae, discs, paraspinal muscles.	Dice ≥0.77; biomechanical models and morphological metrics are highly correlated (R ≥0.69) with human annotations.	Deep learning segmentation enables automated biomechanical modeling for chronic low back pain, scalable for large cohorts.
Xiao et al. (2025)	3D U-net segmentation + automated landmark identification/statistical shape models (SSM) for femoral neck anteversion (FNA).	The computerized method had a DSC of 0.989, strong correlation (R >0.9) with manual methods, and high intra-/inter-observer reliability.	Automated 3D CT FNA measurement is reliable and can support preoperative planning and robotic orthopedic surgery.

Conclusion

Artificial intelligence is reshaping preclinical and translational orthopaedic research, with applications spanning biomechanics, imaging, tissue engineering, drug discovery, genomics, and prosthetic control. AI enhances efficiency, accuracy, and personalization—examples include automated imaging analysis, synthetic data generation, scaffold optimization, and fracture risk prediction. However, the methodological quality of the studies is limited, as many studies faced data scarcity, validation issues, and ethical concerns, thereby undermining the generalizability of the findings. The lack of standardized performance metrics highlights a gap in reproducibility and external validation in AI applications. While other challenges remain, the integration of multimodal data and advanced learning methods offers a path forward. Rather than replacing traditional methods, AI serves as a powerful complement, accelerating translation from laboratory research to patient-specific clinical care. AI is transforming orthopaedics by enhancing imaging, drug discovery, implant planning, and regenerative therapies in both preclinical and translational domains. The field is advancing rapidly, but clinical translation requires better validation, regulation, reproducibility, and explainability.